Yolo Series Network Training data Preparation Tool-yolo_mark

Project Introduction

Yolo_mark is a test task data set making tool, after the completion of the data format is not VOC or Coco data format, from its name can be seen, it is specifically for the Yolo series of network training to prepare data, Yolo this is very willful, It does not use any of the existing deep learning frameworks to implement his code, but instead writes a pure C lightweight framework-darknet, so its training data is not prepared in the same format as a standard open source dataset. Yolo_mark is specifically prepared to prepare Yolo for training data, here is its GitHub address.
The project supports Windows and Linux two systems, depending on the OpenCV library, 2.X or 3.X can be. If it is windows, you need VS2013 or VS2015. Use

Below is a description of how Yolo_mark is used under Windows:
The. sln file is already available in the project, and after you open the solution using VS2013 or VS2015, the project is compiled under the x64 and release platforms, and it is important to note that we need to configure the OpenCV first, and this process is not detailed.
Successful compilation will generate an. exe program in the Release folder under x64 and then run the generated EXE program through the yolo_mark.cmd command-line file.
The effect is this, a very simple GUI interface:

Yolo_mark is still well implemented, and the key question is how the data is organized, as it is used in Yolo's network training. Data Organization

The Data folder in the project holds the required and labeled results:
The. Data file defines the number of categories of target detections, the list of TXT files for the training data and the test data, and the names of each category (it is obtained from the. names) file:

Classes= 1
Train  = data/train.txt
valid  = data/valid.txt
names = data/obj.names
backup = backup/

The name of each category is defined in the. names file, each category has a separate row and the number of rows corresponds to the name one by one, such as single-class pedestrian detection, with only one row in the. Names File:

Human

At this point, Human is ID 0.
The Train.txt file is a list of training data that is automatically captured by the data stored in the IMG folder:

Data/img/0.jpg
data/img/1.jpg
data/img/2.jpg
data/img/3.jpg
data/img/4.jpg
data/img/5.jpg
data/img/6.jpg
data/img/7.jpg

Finally, the IMG folder, before the beginning of labeling, the IMG folder only the image data that needs to be labeled (requires. jpg file), a picture callout is completed, the image will be generated a. txt file with the same name, which is stored in the Bbox information:

A. txt file, each line is a target information, which means that there are several rows of data, the image is marked with a few targets, it is based on "id x y W h" in the form of storage, where "x y W H" is normalized after. For example, the result in a 0.txt file:

0 0.341797 0.547917 0.049219 0.118056
0 0.731250 0.581944 0.050000 0.225000